2334 Background Paper Colin Thirtle Agricultural Productivity Poverty Developing Countries

1
AGRICULTURAL PRODUCTIVITY AND POVERTY IN DEVELOPING COUNTRIES
Extensions to DFID Report No.7946
Lin Lin*, Victoria McKenzie*, Jenifer Piesse** and Colin Thirtle***

*
Environmental Policy and Management Group
T.H.Huxley School of Environment, Earth Sciences and Engineering
Imperial College of Science, Technology and Medicine, London, SW 7 2BP
**
School of Management and Organizational Psychology
Birkbeck College, University of London, London, WC 1E 7HX
and University of Stellenbosch, Stellenbosch, 7602, Republic of South Africa.
***
Environmental Policy and Management Group
T.H.Huxley School of Environment, Earth Sciences and Engineering
Imperial College of Science, Technology and Medicine, London, SW 7 2BP
and University of Pretoria, Pretoria, 0002, Republic of South Africa
OUTLINE
EXECUTIVE SUMMARY
1) INTRODUCTION
2) METHODOLOGY AND DATA
3) RESULTS: EFFECTS OF AGRICULTURAL PRODUCTIVITY GROWTH ON POVERTY AND NUTRITION
3.1 Regressions Using the Cross Section of the Less than $1 Per Day Poverty Index
3.2 Regressions Using Pooled Data for the Less than $1 Per Day Poverty Index
3.3 Regressions Using The Human Development Index
3.4.1 Regressions Using Nutrition Indicators
4) DOES PRODUCTIVITY GROWTH IN INDUSTRY AND SERVICES REDUCE POVERTY?
5) EXPLAINING INEQUALITY AND IMPROVING THE SPECIFICATION OF THE MODEL
5.1 Single equation approach

5.2 Explaining inequality
5.3 Single equations using residuals and recursive equation models
6) R&D COSTS OF YIELD INCREASES TO REDUCE POVERTY & INCREASE GDP PER CAPITA
6.1 Calculating the Rate of return to Investment in R&D
REFERENCES
1
We thank Lawrence Haddad of IFPRI, Jonathan Kydd and Peter Dorward of Imperial College at Wye, for help and advice
and especially Gavin McGillivray of DFID for the original suggestion to investigate the relationship between productivity
and poverty in this manner. Michael Lipton suggested that we should also look at the effect of industrial productivity
growth.
EXECUTIVE SUMMARY
This paper investigates the empirical relationships between agricultural productivity growth, poverty
reduction, nutritional improvement, inequality and GDP per capita growth in some detail. The
empirical estimates of the relationship between labour productivity in agriculture and poverty
reduction appears to be particularly robust. For all the specifications of the model and for all the
different samples, labour productivity in agriculture is found to be a powerful and always significant
cause of poverty reduction.
The study begins with simple models, which concentrate on the explaining poverty measures
with just the key variables, which are the two components of agricultural labour productivity (value
added per unit of land, or yield and the land labour ratio) and the Gini coefficient, which measures
inequality. For the latest cross section of 40 counties, from the 2001 World Development report, the
two agricultural productivity variables alone explain 63% of the variance in the percentage of the
population living on less than $1 per day. The poverty elasticity of the yield is -0.91and is significant
at the 99% level of confidence. For the 19 African countries alone, 47% of the variance is explained,
the significance level is the same despite the small sample and the elasticity is slightly higher at –0.96.
This means that a 1% increase in yields decreases the percentage of the population living on less than
$1 per day by 0.96%.
For a larger pooled sample, in which there are 109 observations on 58 countries, where there
are two or even three observations for some countries, at different points in time, the yield elasticity
falls to 0.621, but the high level of significance is maintained and 50% of the variance is explained.
The double counting of some countries could be the cause of the difference, but it is more likely that
including the Gini coefficient reduces the elasticity. The Gini measure of inequality has a poverty
elasticity of 2.2 and is highly significant. Thus, a 1% increase in inequality increases the percentage
of the population living on less than $1 per day by 2.2%, which is clear evidence of the adverse effects
of increases in inequality. For the African countries alone, the sample is increased to 38 and although
the Gini coefficient has a similar impact, the poverty elasticity of the yield rises to –1.02%, with 55%
of the variance explained.
The $1 per day poverty measures begin only in 1985, so to cover the green revolution period,
this is replaced by the Human Development Index. The agricultural productivity variables remain
significant at high levels of confidence, for both the full sample of 280 observations on 174 countries
and for the smaller African sample of 78 observations. The yield elasticity is 0.12 for the full sample,
with 76% of the variance explained and 0.094 for the Africa sample, with an explanatory power of
48%. Thus, a 1% increase in yields increases the value of the index by 0.12% for the full ample and
by 0.094% for the African countries. Since the units of measurement are totally different, these
results are not comparable to those for the $1 per day poverty measure and they are not so easy to
interpret, but they do confirm the power and robustness of the agricultural productivity variables.
The final applications of this simple model are to two nutrition variables, which are the per
capita dietary energy supply and the percentage of children under five years old that are underweight.
For the full sample of 109 observations, the yield elasticity with respect to energy supply is 0.05,
which indicates that a 1% increase in yields increases the per capita dietary energy supply by only
0.05%. Again, the result is not comparable to the $1 per day poverty estimates, but also the
explanatory power is only 25%, so these two results suggest that nutritional improvement depends on
many other variables and is more difficult to achieve. For the Africa sample the yield elasticity is
0.09, which is almost twice as great, but again only 25% of the variance is explained.
The yield elasticity with respect to the percentage of children that are underweight is
comparable (as it is again a reduction measured in percent) and it is –0.42, which means that a 1%
increase in yields reduces child malnutrition by 0.42%. This is more encouraging, but again only 31%
2
of the variance is explained. For the African sample the yield elasticity falls to –0.27 and only 28% of
the variance is explained. Thus, the yield elasticity with respect to calories is far higher for Africa, but
it seems to be more difficult to improve the diet of children.
These results confirm the predominant view in the literature surveyed in Thirtle et al. (2001),
that agricultural productivity growth can be expected to have an impact on poverty. However, it is fair
to ask how great this is relative to other improvements in economic performance. Thus, the impact of
labour productivity growth in agriculture is first compared with the effects of productivity growth in
industry and services. These results need to be treated with some caution, because there was even
more missing data for industry and services than for agriculture, but regardless of the samples used,
agricultural labour productivity had a poverty elasticity of -0.63, which was significant at least the
95% confidence level, whereas labour productivity in industry and services has no significant impact
on poverty reduction.
Attempts to explain inequality were less successful. The yield and land labour ratios had
significant, but small (-0.05 and 0.06 respectively), elasticities and the only other variables which
consistently explained the Gini coefficients were the percentage of the population that were rural and
the rate of population growth. The elasticity for the percentage of the population that was rural was –
0.20, which indicates more rural societies are less unequal, but this may be because they are poorer
and thus have less differentiation. The elasticity of population growth was 0.19, meaning that 1%
greater population growth increases inequality by 0.2%. This confirms the notion that rapid
population growth tends to increase inequality.
The danger with these simple models is that they may be mis-specified, in the sense of
omitting important explanatory variables, so that the poverty elasticity of the yield variable is biased
upwards, or possibly even downwards. To overcome this problem, data on all the available variables
was collected and then tested by looking for correlations with the poverty and inequality measures.
All the combinations of significant explanatory variables were then used in addition to the agricultural
productivity variables and tests were used to determine which variables improved the models. Gross
fixed investment (GFI), which includes infrastructure investment, was found to be the most
consistently significant variable. This is again in keeping with the findings in the literature: see Fan,
Hazell and Thorat, 1999, for example. The elasticity of GFI in the full sample was –0.93 and the
poverty elasticity of yields was reduced to –0.56, so it does seem that infrastructure is also crucial to
poverty reduction. For the Africa sample the yield elasticity was far higher at –0.95 and the GFI
elasticity was lower at –0.48, which suggests that infrastructure is so poor in the African sample that it
has had less effect. Open economies also appear to have less poverty, as the percentage of GDP that
was traded had an elasticity of –0.26 and GDP growth has positive effects, with an elasticity of –0.16.
The point on the effect of missing variables biasing the yield elasticity is bourn out, but the problem is
not severe. The minimum value of the yield elasticity in these models with five explanatory variables
was still –0.54 and the significance levels were maintained.
The difficulties of modelling these relationships become apparent when GDP per
capita is included in the equation as well as agricultural productivity. Then, the yield elasticity falls to
–0.25 and the elasticity of GDP per capita is –0.47. However, this is not an omitted variable problem.
Rather, agricultural productivity is a key variable in explaining differences in GDP per capita. Indeed,
the yield and the land labour ratio, which are the two components of agricultural labour productivity,
alone explain 84% of the variance in GDP per capita. Similarly, these same two variables explain
22% of the variance in equality, as measured by the Gini coefficient. Thus, there is causality running
from agricultural labour productivity to both GDP per capita and the Gini coefficient, so these three
variables should not appear in the same equation.
There are two solutions to this problem. If the single equation approach is to be maintained,
GDP per capita can be regressed on the agricultural productivity variables and the residual from this
equation, which may be termed “agricultural productivity-free GDP per capita can be used instead of
the original variable. Similarly, the Gini coefficient is regressed on the agricultural productivity
3
variables and the residual used instead of the original Gini. If this is done, the yield elasticity remains
close to its original value, at –0.65, the Gini elasticity is unchanged at 2.21 and the GDP per capita
elasticity rises to –0.59, which is within the range suggested by Hanmer and Naschold (2000). The
other significant variables are the percentage of GDP traded, with an elasticity of –0.35 and GFI,
which has an elasticity of –0.67. These variables explain 60% of the variance in the percentage of the
population living on less than $1 per day.
The preferred alternative is to resort to systems of equations. In the first equation, the
percentage of the population living on less than $1 per day is explained by the Gini coefficient
(elasticity of 1.74), the percentage of GDP traded (elasticity of –0.37), gross fixed investment
(elasticity of –0.75) and GDP per capita (elasticity of –0.76). These variables explain 65% of the
variance in poverty. In the second equation, the Gini coefficient is explained by the percentage of the
population that is rural (elasticity of –0.20) and the rate of population growth (elasticity of 0.33).
These variables explain 42% of the variance in inequality. In the third equation, GDP per capita is
explained by the yield (elasticity of 0.58), the land labour ratio (elasticity of 0.57) and expenditure per
student in primary education (elasticity of –0.14). These variables explain 77% of the variance in
GDP per capita. Thus, it is possible to impose some structure on the complex inter-relationships
between the nine variables in the system and produce credible results.
This is one of two preferred recursive models, but all the results come to the same basic
conclusion, which is that agricultural productivity growth has almost as much effect on poverty
reduction as does GDP per capita growth itself. But, the real point is that the gains from agricultural
productivity growth are not confined to the poor, since the yield also has an elasticity of 0.58 with
respect to growth in GDP per capita. This means that 1% growth in yields both reduces the percentage
of the population lining on less than $1 per day by roughly 0.5% and increases GDP per capita by
almost 0.6%.
The final step is to calculate the cost of achieving a 1% increase in yields. This can be done
within the confines of the recursive system, by adding a fourth equation, in which yield is explained
by agricultural R&D and education. For this R&D data must be collected, so for now rough
calculation will have to suffice. In 1991, the less developed countries spend $8 billion on agricultural
R&D (Alston et al., 2000) and yields have been increasing at 2.4% per annum since the 1960s, so the
cost of a 1% yield increase is £3.33. The GDP of the LDCs in 1990 was about $2,850 billion, so with
a GDP elasticity of 0.56, the payoff to this R&D investment was $16.13 billion. If the lag from
expenditures to gains in output is 5 years, the rate of return is 37%: if it is four years, it rises to 50%,
which is surely an attractive investment. This gives a rough estimate of the rate of return to
agricultural research and the poverty reduction impact is a bonus on top of this figure. The number of
persons living on less than $1 per day was 1,200 million in 1998 and the best estimate of the poverty
elasticity of agricultural research is 0.65. Thus a 1% increase in yields reduces this count by 0.65%,
which is almost 7.8 million. Viewed in this way, agricultural research may well be a useful and cost
effective instrument for reducing poverty, but it is a pretty blunt one as many have argued (Alston et
al., 1995). It needs to be sharpened by aiming it at the poorest segment of the population.
1 INTRODUCTION
4
The previous report on this subject (Thirtle et al, 2001) reviewed the literature on the role of
agricultural productivity in alleviating poverty in developing countries and presented some limited
empirical results suggesting that there are significant relationships between productivity growth and
both poverty and nutrition. The major finding was that the empirical estimates of this relationship
appear to be robust. Regardless of differences in data and formulation, the results showed that a 1%
increase in yields leads to a reduction in the percentage of people living on less than $1 per day of
between 0.6% and 1.2%.
This study extends the previous work in several respects.2 First, the sample is broken up to
show that the relationships hold for Sub-Saharan Africa, which is perhaps the region of greatest
interest. Second, the effects of inequality on the productivity-poverty relationship are further
investigated and the causes of inequality are also considered.3 Third, although the relationship between
agricultural productivity and poverty appeared to be robust, there are many other omitted variables
that may affect the relationship, so these are added to the model. Fourth, adding variables to a single
equation is of limited use, since many of the relationships are collinear and the variables tend to cancel
each other out. Also there is genuine simultaneity between variables, so it is necessary to resort to
multiple equation models to improve explanatory power. Fifth, although agricultural productivity has
a significant effect on poverty reduction, it is fair to ask how the impact compares with other options.
So, in this study the impact of productivity growth in industry and services is compared with the
results for agriculture. Last, the link between agricultural R&D and productivity growth is
investigated, with a view to estimating the costs of achieving poverty reduction by means of increasing
agricultural productivity. Also, the results show that agricultural productivity growth increases GDP
per capita as well as reducing poverty: it is not only the poor that gain.
The paper proceeds as follows. The next section explains the basic methodology and outlines
the data and its sources, before presenting an exploratory correlation analysis to determine the
variables, which appear to be important in explaining poverty and inequality. Section three describes
the basic regression models and the results they generate for the full sample and for the African
countries alone. Section four contrasts the impacts of agricultural productivity growth with the effects
of labour productivity growth in industry and services. Section five considers the ways in which the
specification of the model can be improved, by removing the interactions between the variables in the
single equation model and resorting to recursive models. Last, the R&D costs of reducing poverty and
increasing GDP per capita by means of yield increases are estimated and the rate of return to
agricultural R&D is estimated, before the conclusion summarises the results.
2 METHODOLOGY AND DATA
2.1 Basic Regression Models

The basic approach is ordinary least squares regressions, with the poverty index as the
dependent variable. The explanatory variables are alternative measures of agricultural productivity and
other variables that may be expected to have an impact on poverty. The intention is to estimate poverty
reduction elasticities with respect to agricultural productivity, similar to the poverty elasticities with
respect to growth (Chen and Ravallion, 2000 Hanmer and Naschold, 2000, Ravallion, 1997, Ravallion and
Datt, 1999.). Since there is no well-defined economic theory of poverty alleviation, the regression are
somewhat ad hoc and the possibility of misspecification due to variable omission is obviously a danger.
2
Many of the results from the empirical section of the previous paper are reported here, as they are needed to
allow comparisons with the new estimates. Thus, this paper can stand alone, but the previous paper literature
review is needed for the literature review, which led to this empirical investigation.
3
Maxwell (2001) discusses the reinstitution of inequality on the poverty agenda, for instance in the World Bank
World Development and refers to work by Howard White suggesting international targets for inequality
reduction.
5
We exploit the link between labour and land productivity used by Hayami and Ruttan (1986),
which is stated here in value added terms as an identity:
VALUE ADDED VALUE ADDED LAND

≡ × (1)
PER UNIT LABOUR LAND PER UNIT LABOUR
Value added is net of the costs of intermediate inputs, which will remove the cost effects of intensification
using increasing amounts of modern inputs. This should make this measure closer to total factor
productivity (TFP) than the total output-based measure. In this way labour productivity can be decomposed
into the product of two components: land productivity, or yield, and the land labour ratio, which can be
viewed as an indicator of a country’s resource endowment. Thus, the yield contribution to labour
productivity can be separated from the relative scarcity of land, which it is not possible to change. This is
important because a country such as the USA has several hundred times as much land per unit of labour as
a land scarce country, such as Bangladesh, and will have far higher labour productivity as a result.
Thus, the poverty indicator is explained by the productivity indices for land and for labour, by
means of cross sectional regressions, in models 1 and 2. Then, the decomposition is used, in equation 3,
making the land labour ratio and land productivity the independent variables. In these three basic
models all the variables are expressed in logarithms, so that the coefficients can be interpreted as
elasticities, the Xi are the other relevant variables (which vary, but usually include the Gini index) and ε
is the error term.
Value added 
Model 1: ln Poverty Index = α + β ln   + φi ln X i + ε (2)
 Land
 Value added 
Model 2 : ln Poverty Index = α + β ln  + φi ln X i + ε (3)
 Labour 
Value added   Land 
Model 3 : ln Poverty Index = α + β ln   + δ ln  + φi ln X + ε (4)
 Land   Labour 
These three models form the starting point of the analysis. The more complex models developed
later, such as the recursive estimation of the poverty equation and an inequality equation, are described
in section five, when they are applied.
2.2 Data
Dependent Variables: Poverty and Nutrition Indicators
The first poverty indicator used was the percentage of the population living on less than $1 per
day, taken from the World Development Report 2000/2001. This gives a simple cross section, with a
minimum of 40 countries, which are listed later.
To increase the size of the data set it is necessary to pool the results of poverty surveys, so that
for some countries there are two or even three observations at different points in time. This may bias the
results, but it does increase the sample size to 109, which is far more satisfactory. The observations are
for 58 developing countries and range in time from 1985 to 1995. They are from the World Bank and
Chen, Datt and Ravallion, (1994) and were used by Hanmer and Naschold (2000), who we thank for
making these data available.
The earliest observations for the $1 per day poverty index are for 1985, which does not allow
coverage back to the green revolution period. The available index that does go back further, at least to
the final stages of the green revolution, is the Human Development Index (HDI), from the UNDP. The
HDI is a composite index of development. The three most crucial components of the HDI are measures
of longevity, education and income and it may serve as a reasonable poverty index. Educational
attainment, income, and life expectancy are all associated with poverty. Thus, the HDI is used as a
6
second poverty indicator, although is obviously not as satisfactory. The HDI covers 174 countries for
1975, 1980, 1985, 1990, 1998, giving a total number of observations without missing data of 280.
The nutrition indicators are from Lawrence Haddad of IFPRI. There are 181 mixed cross section
and time series observations, but again this reduces to 109 observations. The dependent variables are per
capita dietary energy supply and the percentage of children under five who are underweight.
In later models the Gini coefficient is added as a dependent variable, to give a two equation
recursive model. These data are also from Hanmer and Naschold (2000).
Explanatory variables
The data for all the independent variables were obtained from the World Development Indicators
(2000) CD. The variables are listed in Table 1, which also reports the correlation of the variables with
the $1 per day poverty measure and the Gini coefficients. The correlation coefficients provide some
guidance as to which variables are likely to have explanatory power in the regression models. These
significant variables are picked out in red. For the percentage of the population living on less than $1
per day poverty index, a negative sign indicates that the variable is likely to be poverty reducing and a
positive sign the opposite. The range of the Gini is from zero, which indicates perfect equality, to unity,
which would be complete inequality, so again a negative sign suggests that the variable may decrease
inequality. Thus, the Gini coefficient has a detrimental effect on poverty reduction, because the positive
coefficient means that greater inequality correlates with a higher percentage of the population living on
less than $1 per day, which is the result suggested by the literature (see Hanmer and Naschold, 2000, for
example). Conversely, the three agricultural productivity variables from equations (1)-(3) are all
significantly poverty reducing, whereas productivity growth in industry and services has no significant
effect on poverty, but does decrease inequality, as does land productivity.
The other explanatory variables are those available that seemed most likely to affect poverty and
inequality, some of which appear in the literature. The combination of the signs on $1 a day poverty
and inequality correlations is sometimes interesting. Of the other agricultural variables, the rural
population as a percentage of the total and agriculture’s share in value added are both positively
correlated with poverty, indicating that counties with large agricultural sectors tend to be poorer.
However, the negative correlations with the Gini coefficient show that they also lessen inequality. The
percentage of the land that is irrigated has no significant effects.
The literature review in Thirtle et al (2001) found that there was evidence that health and
education expenditures may reduce poverty, so literacy and health expenditures were included and
were found to be significantly poverty reducing, according to the correlation coefficients.4 Literacy
also correlates with less inequality. Educational expenditures per student in primary education
correlates negatively with inequality, but has no significant impact on poverty. Educational
expenditures per student in secondary education are not significant and neither of these expenditure
variables is correlated with literacy. Similarly, infrastructure has been shown to be important in the
literature (Fan, Hazell and Thorat, 1999, for example), so gross fixed investment is included to try to
capture infrastructure investment.5 This is also negatively correlated with poverty.
Table 1: Socio-Economic Dataset
4 Note that the variable is illiteracy, which is why the sign appears to be wrong.
5 Gross domestic fixed investment includes land improvements (fences, ditches, drains, and so on); plant,
machinery, and equipment purchases; and the construction of roads, railways, and the like, including
commercial and industrial buildings, offices, schools, hospitals, and private residential dwellings. Data are in
constant 1995 U.S. dollars. For more information, see WDI table 4.10.The World Bank data used to include
government expenditure on transport and communications, which would have been preferable, but this has been
discontinued.
7
Variable Name Correlation Correlation Sample
with $1 per with GINI Size
day poverty Coefficient
1 Dollar per Day (% Total Population)* 1.00** 0.35** 121
* **
Gini Coefficient 0.35 1.00** 121
Agriculture value added per worker (constant 1995 US$) -0.55** 0.12 113
** **
Agriculture value added per hectare of agricultural land -0.30 -0.41
(constant 1995 US$) 109
Agriculture land per worker -0.18** 0.44** 109
Services, value added per worker (constant 1995 US$) -0.05 -0.23** 57
Industry, value added per worker (constant 1995 US$) -0.02 -0.23** 54
** **
Rural population (% of total population) 0.42 -0.17 121
Agriculture, value added (% of GDP) 0.43** -0.16** 119
Land use, irrigated land (% of cropland) -0.12 -0.04 115
Health expenditure per capita, PPP (current international $) -0.44** 0.19 54
** **
Illiteracy rate, adult total (% of people aged 15 and above) 0.50 0.21 115
Expenditure per student, Primary Education (% of GDP) -0.10 -0.26** 61
Expenditure per student, Secondary Education (% of GDP) -0.28 -0.09 28
Gross domestic fixed investment (% of GDP) -0.33** 0.02 118
Trade (% of GDP) -0.32** 0.09 119
Exports of goods and services (% of GDP) -0.42** 0.06 119
General government consumption (% of GDP) -0.07 0.13 119
State-owned enterprises, economic activity (% of GDP) -0.18** -0.03 52
**
Tax revenue (% of GDP) -0.47 0.05 87
GDP per capita, (current international PPP $) -0.61** -0.01 117
GDP growth (annual %) -0.02 -0.013 90
Population growth (annual %) 0.29** 0.35** 114
Unemployment, total (% of total labour force) -0.14 0.02 45
Bread and cereals price in PPP terms (U.S. price = 100) -- -- 0
Inflation, consumer prices (annual %) 0.04 0.06 108
* From Hanmer and Naschold (2000) ** significant at the 5% level, two-tailed test.
The liberalisation literature argues that openness is important and both the percent of GDP
traded and exports as a percentage of GDP are negatively correlated with poverty. The damaging effects
of overly-large public sectors is also a part of the liberalisation creed, but government consumption is
not significant and the percentage of GDP accounted for by state owner enterprises is negatively
correlated with poverty. This may be a function of the sample, but perhaps the countries with larger
government sectors do tend to be more egalitarian (there are no formerly communist countries in this
sample, which could have been the cause of this result). Inflation is also slated as a bad, but has no
discernable effect on poverty or inequality. However, tax revenue also correlates with less poverty, so at
least some taxation would seem to be re-distributive.
The key variable in the literature on the poverty-reducing effect of growth (surveyed by Hanmer
and Naschold, 2000) is GDP per capita, which also shows a strong negative correlation with poverty, but
not with inequality and the GDP growth rate is not significant. Population growth is positively correlated
with both poverty and inequality, indicating that countries with more rapid population growth do
experience difficulties, but inflation has no discernable impact.
These results are summarised in Tables 2 and 3, which list variables in descending order of
correlation with first the $1 per day poverty index and then the Gini coefficient. Thus, Table 2 first lists
the poverty reducing variables, showing that GDP per capita has the highest correlation with poverty
reduction, followed by agricultural labour productivity. The correlation coefficients are bivariate
relationships, which give some indication of the possible impacts, but they take no account of the
multiple causes of poverty reduction or the interactions between the variables and correlation implies
8
nothing about the direction of causality. Thus, although the evidence from the literature suggests that
GDP per capita growth reduces poverty, it is also true that having more poor people in the population
reduces GDP per capita. Despite these caveats, it is worth noting that the effect of agricultural labour
productivity has almost as much impact as growth in GDP per capita. This is consistent with the lack of
impacts on poverty of productivity growth in industry and services. Also, the costs of increasing
agricultural productivity growth are very low, relative to the cost of an equal increase in GDP per capita.
It would be surprising that productivity growth in a single sector can have such a big impact, were it not
for the extensive literature, which almost all points to the link between agricultural growth and poverty
reduction. The actual impacts will be better established with multivariate regressions, in the next
section.
The last five variables in Table 2 are those that appear to increase the percentage of the
population living on less than $1 per day. Again, these results are well supported by the literature, which
has always argued for basic education as a means of improving the position of the poor. Table 3 repeats
this summary exercise for the Gini measure of inequality, showing that increasing yields has the greatest
impact on inequality reduction. Although there was no direct poverty impact of labour productivity in
industry and services, they must have an indirect effect, by way of reducing inequality. Last, the
inequality-increasing effect of more land per worker is difficult to interpret. It could well result from the
Asian countries with more population pressure on the land have less poor people than the African
countries, which have far more land, of much worse quality.
To conclude the data analysis in this section, Tables 2 and 3 also show the summary statistics for
all the variables in Table 1, for the full sample and for the African countries alone, so that the two can be
compared. The mean values in the first part of Table 2 show that the Africa sample has substantially
lower GDP per capita and far lower agricultural land and labour productivity. The higher land labour
ratio is not quality adjusted: if the FAO data on land potential were taken into account this difference
could even be reversed, as these data show that Kenya is at least as land-scarce as the heavily populated
Asian countries, once land quality and rainfall differences are included. Africa has a greater proportion
of GDP accounted for by state-owned enterprises and this variable was negatively correlated with
poverty.
The lower part of Table 2 shows that literacy rates in Africa are far lower than for the full
sample and that agriculture has a higher share of GDP in Africa, which is not surprising, given the much
higher proportion of the population that are rural. The Gini coefficients show that there is not much
difference in inequality levels, but if anything this Africa sample has greater inequality. Finally,
population growth is greater for the Africa sub-sample.
The top part of Table 3 repeats some of the Table 2 results, but also shows that Africa actually
spends more per student on primary education, despite, or perhaps because of the higher level of
illiteracy. Labour productivity in industry for the Africa sample is only 53% of the level in the full
sample and in services the difference is even more pronounced, at only 35%. The lower part of the Table
adds no new information.
This completes the exploratory analysis of the data and suggests that the agricultural
productivity variables are strongly correlated with poverty and inequality. We now use the information
gleaned above in constructing regression models to explain poverty and inequality reduction.
9
Table 2: Variables Correlated with $1 Day Poverty
Poverty Reducing Correlation Mean Standard Minimum Maximum

Deviation
All Africa All Africa All Africa All Africa
GDP per capita, (current international PPP $) -0.61** 2799.8 1579.0 2096.4 1591.9 335.6 335.6 10948.0 8029.3
Agriculture value added per worker (constant -0.55** 1441.8 622.1 1363.8 754.8 164.4 164.4 5319.5 3388.8
1995 US$)
Tax revenue (% of GDP) -0.47** 17.6 19.0 8.6 8.0 3.6 8.5 44.7 38.6
Health expenditure per capita, PPP (current -0.44** 187.9 147.2 154.3 180.2 10.0 10.0 784.0 571.0
international $)
Exports of goods and services (% of GDP) -0.42** 27.2 24.8 16.4 14.3 5.5 5.6 89.4 62.0
Gross domestic fixed investment (% of GDP) -0.33** 20.6 20.1 8.6 12.4 5.8 5.8 71.4 71.4
Trade (% of GDP) -0.32** 60.9 61.3 34.6 30.6 13.9 19.7 182.7 151.4
Agriculture value added per hectare of -0.30** 317.3 104.0 323.2 102.5 4.3 4.3 1444 432.9
agricultural land (constant 1995 US$)
Agriculture land per worker -0.18** 10.7 16.6 17.6 27.6 0.29 0.6 102.5 102.5
State-owned enterprises, economic activity (% -0.18** 10.5 13.8 8.2 11.0 0.6 5.3 35 35
of GDP)
Poverty Increasing
Illiteracy rate, adult total (% of people aged 15 0.50** 32.4 48.4 23.1 19.2 0.4 17.4 88.0 88.0
and above)
Agriculture, value added (% of GDP) 0.43** 22.2 29.6 13.9 15.7 4.2 4.2 60.8 60.8
Rural population (% of total population) 0.42** 57.3 90.6 20.7 13.7 15.7 42.1 95.0 95
Gini Coefficient* 0.35** 0.433 0.467 0.111 0.102 0.205 0.289 0.634 0.623
Population growth (annual %) 0.29** 2.4 2.9 0.9 0.5 0.1 1.8 6.6 3.7
** Significant at 5% level.
Table 3: Variables Correlated with GINI Coefficient
Mean Standard Minimum Maximum
Inequality Reducing Coefficient Deviation
All Africa All Africa All Africa All Africa

Agriculture value added per hectare of -0.41** 317.3 104.0 323.2 102.5 4.3 4.3 1444 432.9
agricultural land (constant 1995 US$)
Expenditure on Primary Education (% of -0.26** 13.4 17.4 8.5 9.4 2.5 6.3 47.4 47.4
GDP)
Services, value added per worker (constant -0.23** 13641.1 4719.7 20413.4 5518.0 459.4 459.4 91755.2 19898.3
1995 US$)
Industry, value added per worker (constant -0.23** 14368.3 7678.0 21984.4 9446.8 374.7 374.7 125846.1 37127.2
1995 US$)
Rural population (% of total population) -0.17** 57.3 90.6 20.7 13.7 15.7 42.1 95.0 95
Agriculture, value added (% of GDP) -0.16** 22.2 29.6 13.9 15.7 4.2 4.2 60.8 60.8
Inequality Increasing
Agriculture land per worker 0.44** 10.7 16.6 17.6 27.6 0.29 0.6 102.5 102.5
Population growth (annual %) 0.35** 2.4 2.9 0.9 0.5 0.1 1.8 6.6 3.7
Illiteracy rate, adult total (% of people aged 15 0.21** 32.4 48.4 23.1 19.2 0.4 17.4 88.0 88.0
and above)
*From Hanmer and Naschold (2000)
** Significant at 5% level.
11
3 REGRESSION ANALYSIS AND RESULTS
3.1 Regressions of the Cross Section of $1 per Day Poverty from WDR 2000
The analysis begins by fitting equations (1)-(3) with the available cross section of $1 per day
poverty percentages as the dependent variable. The results are reported in Table 4. Model 1 has only 40
observations because the yield data ends at 1995 and many of the 72 poverty estimates are for later
dates. However, the 40 countries are a reasonable sample of the developing world. They are Algeria,
Botswana, Bulgaria, Burkina Faso, Central African Republic, Chile, Cote d'Ivoire, Ecuador, Egypt,
Estonia, Guatemala, Kenya, Korea, Lesotho, Madagascar, Mali, Mauritania, Mexico, Mongolia,
Morocco, Namibia, Nepal, Niger, Paraguay, Poland, Portugal, Romania, Rwanda, Senegal, Sierra Leone,
Slovenia, South Africa, Sri Lanka, Tanzania, Tunisia, Turkey, Uganda, Uruguay, Uzbekistan and
Zimbabwe. The greatest weakness is that the largest countries of Asia, such as China and India, are
missing, so a huge proportion of the world’s poor people are not included. Also, some would omit the
central and eastern European countries.
The adjusted R2 of 0.20, in model 1of Table 4, means that yields explains only 20% of the
variance in poverty, which is not satisfactory, since it suggests that other, omitted variables explain the
majority of the differences, but the productivity measure is significantly different from zero at a high
level of confidence. The poverty elasticity of – 0.37 means firstly that higher yields result in lower
percentages of the population living in poverty and secondly that a 1% increase in yields reduces the
percentage of the populations living on less than $1 per day by 0.37%.
Table 4: Dependent Variable is % of Population with Less than $1 per Day: Cross Section
Explanatory Expected Sign Estimated Coefficients

Variables
Model 1 Model 2 Model 3
VA/LAND Negative -0.37** -0.91**
VA/LABOUR Negative -0.83**
LAND/LABOUR Negative -0.819**
** **
Constant 4.26 8.06 8.48**
R square 0.20 0.506 0.625
F Test 13.35** 13.35** 53.42**
Sample Size 40 66 40
** significant at the 1% level, two-tailed test.
Model 2 explains just over 50% of the variance, which is far better and the poverty elasticity,
which is again highly significant, rises to – 0.83, so a 1% improvement in labour productivity reduces
the poverty count by 0.83%. The sample increases to 66, but the problem with this model is that the
effect could all be coming from the land-labour ratio component of the labour productivity index.
Thus, following these preliminary tests, Model 3 separates the two terms. The model explains
62% of the variance in poverty and the large increase in the F statistic indicates that it is statistically
preferred to the two previous attempts. A 1% increase in the land labour ratio reduces poverty by
0.82%, which is surprisingly low relative to the effect of the land productivity term, which indicates that
a 1% improvement in yields decreases the percentage of the population living on less than $1 per
day by 0.91%. Again, the variables are highly significant and this is the preferred model. The result
can be developed further if an elasticity can be calculated to link R&D expenditure to yield gains.
Then, the cost of generating a 1% decrease in poverty could be calculated. Since R&D expenditures are
quite modest, our expectation is that this could look like a very cost effective means of reducing poverty.
The exercise is repeated for the 19 African countries, since Africa is regarded as a major
problem area. Although the sample size is barely sufficient, the results are still robust. Table 5 shows
that yield alone has no explanatory power, but the poverty elasticity of labour productivity changes very
little and once this is decomposed into its two elements in model 3, both are significant. Indeed, the
poverty elasticity of land productivity rises from –0.91 to –0.96 and the land/labour ratio adjustment also
increases slightly. The land/labour ratio is playing an important role in making the yield variable
significant for the Africa sample, by controlling for land quality. The countries with poor soil and water
have far more land per unit of labour, so this variable is quality adjusting the data to allow the yield
effect to show through. Whilst this is an encouraging result, it is also true that a 1% increase in yields is
probably harder to achieve in Africa.
Table 5: Dependent Variable is % of Population with Less than $1 per Day: Africa only
Variables
VA/LAND Negative -0.022 -0.96**
**
VA/LABOUR Negative -0.87
Constant 3.19** 8.46** 8.92**
R square 0.0005 0.42 0.47
F Test 0.01 14.39** 7.11**
3.2 Regressions using Pooled Data on $1 per Day Poverty
Whereas the WDR data has only single observations for each country, the data used by
Hanmer and Naschold (2000) has scattered observations from 1985 to 1995 for 58 countries, which
increased sample size to 109 observations.6 Models 1 to 3 are the same as in the previous section and the
results are reported in Table 6.
Table 6: Dependent Variable is % of population with less than $1 per day: Pooled Sample
Variables Expected Sign Estimated Coefficients
Model 1 Model 2 Model 3 Model 4
VA/LAND Negative -0.299** -0.72** -0.621*
**
LAND/LABOUR Negative -0.605** -0.742*
GINI Positive 2.153*
YEAR DUMMY -0.014 0.117 0.1066 0.185
Constant 4.498* 7.177* 7.616* -0.776
R square 0.088 0.3095 0.328 0.50
F-statistic 6.44** 20.2** 14.79** 15.66**
Sample Size 109 113 109 109
* significant at the 5% level, two tailed test. ** significant at the 1% level, two tailed test.
In model 1, the 1$ a day poverty indicator is regressed on land productivity and only 9% of the
variance of poverty is explained. Dummy variables were included to allow for the different time periods,
but the coefficients were not significant. Again, the productivity measure is significantly different from
zero at a high level of confidence. From the poverty elasticity we can infer that an increase of 1% in
labour productivity would bring about a 0.3% decrease in the poverty headcount index
In model 2, the poverty indicator is regressed on labour productivity, which explains 30% of
the variance and gives a highly significant elasticity of 0.63. Model 3, where the $1 per day poverty
indicator is regressed against both land productivity and the labour land ratio, explains over 32% of the
variance in poverty. Both components, land productivity and the land labour ratio, are significant at the
5% level. The poverty elasticities indicate that if land productivity were to increase by 1% there would
be a 0.72% reduction in the percentage of the population living on less than $1 per day, whilst if the land
6
We thank them for making these data available.
13
labour ratio were to increase by 1%, this would bring about 0.6% decrease in the percentage of people
living on less than $1 per day.
Model 4 adds the Gini coefficient, which is an index of inequality, varying from zero, which is
perfect equality to unity, which would be complete inequality. Throughout the literature review, it was
suggested that greater inequality prevented growth from reducing poverty. The adjusted R2 in Model 4
rises to 50% and all three variables are statistically significant at the 5% level. The results infer that a
1% increase in land productivity would reduce the poverty headcount index by 0.62 and that a 1%
increase in the land/labour ratio would reduce poverty the poverty headcount index by 0.62%. However,
the most striking effect is that if there were a 1% decrease in the Gini index, there would be a 2.19%
decrease in the poverty headcount index. Thus, the larger sample gives a very similar result to the first
cross section regressions. The poverty reduction from a 1% increase in yields still appears to be between
0.6% and 0.7% and the relationship is again highly significant.
Again, the exercise is repeated for the African countries in the sample, which now gives a more
reasonable count of 38 observations. Again, the effect of yield alone is not significant, due to the huge
differences in land quality and water availability in Africa, but the poverty elasticity of labour
productivity rises from –0.629 in the full sample to almost exactly –1.0 for the African countries. Thus,
productivity growth has a greater impact in the more predominantly agricultural countries of Africa and
this effect is just as marked in the preferred models which decompose labour productivity. A one
percent increase in yields gives a slightly greater than one percent reduction in the percentage of the
population living on less than $1 per day. Note too that although the Gini coefficient is much the same
size it is not significant. This may be due to the smaller sample size, but is also affected by inter-
relationships with the agricultural productivity variables, which will be pursued later.
Table 7: Dependent Variable is % of population with less than $1 per day: Pooled Africa sample
VA/LAND Negative -0.11 -01.04** -1.02**
**
LAND/LABOUR Negative -0.96** -0.99**
GINI Positive 2.23
YEAR DUMMY -0.33 0.15 0.15 0.13
Constant 3.82** 9.44** 9.59** 8.64
R square 0.03 0.55 0.55 0.55
F-statistic 0.68 8.26** 5.79** 4.3**
Sample Size 38 38 38 38
** significant at the 1% level, two tailed test.
3.3 Regressions using the Human Development Index

The earliest observations for the $1 per day poverty index are for 1985, whereas the HDI goes
back to 1975, so the HDI was used in the place of the $1 per day poverty measure in an attempt to cover
the effects of the green revolution period. Only the preferred model 3 is reported as the preliminary tests
have been repeated several times. The two agricultural productivity variables alone explain 76% of the
variance in the HDI and both are highly significant. Thus, this regression confirms the apparently solid
link between agricultural productivity growth and poverty reduction. Raising yields by 1% increases
the HDI by 0.12%, which is the right direction, but improving the value of a composite index does not
have the obvious and appealing meaning of reducing the $1 per day measure.
Table 8: Explaining the Human Development Index: Full sample

Model 3
VA/LAND Positive 0.1226**
LAND/LABOUR Positive 0.1011**
14
Constant -2.39**
R square 0.759
F-statistic 328.48**
Sample Size 280
For the Africa only sample, the elasticities smaller but are still significant, but only 48% of the
variance is explained. This is a reasonable result in that the green revolution was a predominantly Asian
phenomenon and despite some successes in Africa, agricultural technology has had a lesser impact.
Table 9: Explaining the Human Development Index: Africa sample

Model 3
Constant -1.01**
R square 0.48
F-statistic 20.86**
Sample Size 79
3.4 Regressions using Nutrition Indicators

Here the data are from Lawrence Haddad of IFPRI. There are 181 mixed cross section and time
series observations, from 1971-96, but again this reduces to 109 observations in model 3, which is again
the most successful regression. Dummy variables were again used to deal with the time difference. Two
variables D8089 and D9096 were generated to adjust for the time differences in the data. Thus, there are
two time dummies, for the 1980-1989 and the 1990-1996 respectively, which measure differences
relative to 1971-79.
In the first case, the dependent variable is per capita dietary energy supply and 22% of the
variance is explained by land productivity and the land labour ratio. The elasticities are highly
significant and a 1% increase in land productivity increases the energy supply by 5.3%. This seems
somewhat low, especially relative to the results for the second nutrition variable.
Table 10: Explaining Per Capita Dietary Energy Supply: Full sample
Model 3
DUMMY 8089 0.02
DUMMY 9096 0.016
Constant 7.38**
R square 0.22
F-statistic 9.89**
Sample Size 109
**significant at the 1% level, two tailed test.
For the Africa sample, which was limited to 35 observations, the elasticites are considerably
larger than for the full sample, so improvements in land productivity may have been lower than for Asia,
but their effects on calorie consumption are greater, for this poorer group of countries.
Table 11: Explaining Per Capita Dietary Energy Supply: Africa sample
15
MODEL 3

LAND/LABOUR Positive 0.07*
DUMMY 8089 0.03
DUMMY 9096 -0.04
Constant 7.17**
R square 0.25
F-statistic 2.35*
Sample Size 35
**significant at the 5% level, two tailed test.
*significant at the 10% level, two tailed test.
Table 12 shows 30% of the variance in under-weight five year old children is explained by the
two variables and that a 1% increase in yields decreases the percentage by 0.42%.
Table 12: Explaining the % of Under-weigh Children below 5 years old: Full sample
Model 4
VA/LAND Negative -0.42**
D8089 -0.21
D9096 -0.3
Constant 5.04**
R square 0.31
F-statistic 7.63**
Sample Size 109
For the Africa sample, the results in Table 13 show that the elasticities are still significant, but in
contrast to the calorie consumption results, they are smaller than for the full sample, rather than larger.
This would seem to indicate that although increases in land productivity have a bigger impact on energy
consumption, the share going to mothers and children is lower in Africa.
Table 13: Explaining the % of Under-weigh Children below 5 years old: Africa sample
Model 4
LAND/LABOUR Negative -0.2*
D8089 -0.26*
D9096 0.15
Constant 4.76**
R square 0.28
F-statistic 2.6**
Sample Size 38
* significant at the 5% level, two tailed test.
Summary
Agricultural productivity increases have significant effects on all the poverty and nutrition
measures, both for the full sample and for the African countries alone. There is also clear evidence
that inequality, as measured by the Gini coefficients, has a substantial negative impact on poverty
reduction. However, although these results appear to be robust, before the actual magnitudes can be
16
believed, further work is required. Particularly, Tables 1-3 suggested that there are many other
variables affecting poverty and inequality, so these must be taken into account. Otherwise, variable
omission may bias the elasticities. Specifically if a variable that is positively correlated with the
explanatory variables is omitted, this will tend to bias the elasticities upwards, since the contribution
of the omitted variable will tend to be picked up by those that are included. The reverse will be true if
the omitted variable is negatively correlated with the regressors. However, we first compare the
impacts of agricultural productivity growth with the effects of productivity growth in industry and the
services sector.
4 DOES PRODUCTIVITY GROWTH IN INDUSTRY AND SERVICES REDUCE

POVERTY?
The effect of agricultural productivity on poverty reduction appears to be consistent and
substantial, but is agriculture different? It is possible that productivity growth in industry or services
may have similar effects. The literature review in Thirtle et al (2001) suggested that agricultural
productivity growth should have a greater effect, but this proposition has not yet been tested. Thus, we
now extend the analysis to industry and services.
First, the correlations between productivity in the three sectors are reported in Table 14, which
suggests that the literature is indeed correct.7 The first column shows that the correlation coefficient
between labour productivity in services and the percentage of the population living on less than $1 per
day is only –0.05, which the probability value shows is not significantly different from zero. Similarly,
the correlation coefficient for industry is only –0.015, which is again insignificant. However, the
correlation coefficient for agricultural productivity is –0.55 and this is significant at the 99% confidence
level. This is 11 times greater than the services coefficient and 36 times larger than that for industry.
The other essential finding in the correlation matrix is the correlation between labour productivity in
industry and services, which is 0.888. This is high enough to suggest that if both are included in the
same regression, they will cancel each other out, in the sense of making both elasticities insignificant.
Table 14: Inter-sectoral productivity correlations

$1 per day Poverty VA/Labour (service) VA/Labour (industry)
VA/Labour (service) -0.0524 1

(p-value) 0.6984
VA/Labour (industry) -0.015 0.8879 1

(p-value) 0.9141 0
VA/Labour
(agriculture) -0.5545 0.0329 0.085
(p-value) 0 0.8152 0.5571
Thus, the first regressions fit the productivity variables individually, although this may be a poor
specification. Since bivariate regression is very little different from correlation, it is not surprising that
the elasticities reported in Table 15 are very similar to the correlation coefficients. The differences that
do occur can be attributed to the logarithmic transformation, which makes the coefficients elasticities.
The R2 values for the services and industry regression show that these variables have almost no
explanatory power and this is confirmed by the t statistics, which show that both are insignificant in
explaining the percentage of the population living on less than $1 per day. In contrast, agricultural
labour productivity alone explains 31% of the variance and the poverty elasticity is highly significant.
However, the models are fitted to the available samples, which differ for the three sectors.
7
The data on value added for industry and services has missing data before 1992, which required some
interpolation.
17
Table 15: Dependent Variable is % of Population with Less than $1 per Day: Sectoral
Productivities

Variables
VA/LABOUR (Service) Negative -0.054
VA/LABOUR (Industry) Negative -0.014
VA/LABOUR (Agriculture) Negative -0.63*
** **
Constant 3.25 2.97 7.2**
R square 0.003 0.0002 0.31
F Test 0.75 0.97 6.44**
Since, the test is more reliable if the same sample is used for all three sectoral productivities, this
is the next step and now all three variables are included in a single equation. The results, reported in
Table 16 as model 4 show that the single equations outcomes are changed. The elasticities are of the
same magnitude, but the sign is reversed for industry and only agricultural productivity is significant.
However, the high correlation between labour productivity in services and industry makes this result
suspect, so we next resort to two stage estimation in which the two are not included together. Still, the
results, reported in models 5 and 6 change very little, in that industrial and service sector productivity are
insignificant. These data show that there is little doubt that labour productivity in agriculture has a
significant impact on poverty reduction, whereas productivity growth in industry and services has no
discernable effect.
Table 16: Dependent Variable is % of Population with Less than $1 per Day: Sectoral
productivities

Variables 2-Stage estimation
VA/LABOUR (Service) Negative -0.05 (exogenous) 0.06
VA/LABOUR (Industry) Negative 0.09 0.04 (exogenous)
VA/LABOUR (Agriculture) Negative -0.65* -0.64* -0.64*
Constant 6.9** 6.98 **
6.8**
R square 0.29** 0.29 0.29
F Test 3.29 4.48 3.41**
5 EXPLAINING INEQUALITY & IMPROVING THE MODEL SPECIFICATION
5.1 Single equation approach

The results in the previous two sections are an improvement on the simple correlations of section
2, but as Table 1 showed, there are several other variables, which appear to highly correlated with
poverty and inequality. Prior knowledge suggests that some of these should be included in models that
explain poverty reduction, and misspecification, due to variable omission can bias the estimates of the
elasticities that have been reported. Thus, we now attempt to extend the model to include all the
18
variables that explain poverty and inequality reduction. Given the limited sample size, the dubious
quality of some of the data and the co linearity of the explanatory variables, not too much can be
expected.
Comparisons can be made with other models in which the less than $1 per day poverty index is
used as the dependent variable, especially Hanmer and Naschold (2000), who use the same poverty data.
The models that are reported were chosen on the basis of goodness of fit (the F statistic is preferred to
the R2, although the latter was not ignored), the significance of the variables and particularly analysis of
the residuals. The last takes precedence, since well-behaved, randomly distributed residuals indicate that
the model is free from the econometric problems that can bias the estimates. All the significant variables
from Table 1 were included in these tests, in various combinations, but the number of significant
variables that could be retained in a single equation was extremely limited.
Thus, the only additional variable that improved the model for all three samples was gross fixed
investment, which was included to capture the effects of infrastructure. Table 17 shows that the
preferred model retains the agricultural productivity variables and the Gini, in addition to gross fixed
investment. Model 3 should be compared with model 4 in Table 9, which used the same full sample of
109 observations. Adding the investment variable increases the explanatory power of the model from
50% to 58% and increases the F statistic from 15.66 to 22.31, which is a substantial improvement. The
key finding is that the poverty elasticity of the yield variable is reduced only from –0.62 to –0.56. Gross
fixed investment has a larger elasticity of –0.93, which indicates its importance. The elasticity for the
Gini is far larger, at 2.28, but this variable has a very limited variance, so the actual magnitude could be
misleading.
Table 17: Dependent Variable is % of population with less than $1 per day
Variables
(Non-Africa) (Africa) (Combined)
VA/LAND Negative -0.61*** -0.95*** -0.56***
*** ***
LAND/LABOUR Negative -0.81 -0.92 -0.71***
GINI Positive 2.76*** 0.72 2.28***
*** *
GROSS FIX INV. Negative -1.47 -0.48 -0.93***
**
Constant 1.55 7.74 1.16
R square 0.59*** 0.59** 0.58**
F Test 14.61*** 7.83*** 22.31***
*** significant at the 1% level, two tailed test. ** significant at the 5% level, two-tailed test. * significant at the 10% level,
two-tailed test.
For the non-African countries, the results are essentially similar to those for the full sample, but
the Africa only estimates are more interesting. These results can be compared with model 4, in Table 7,
and now inequality appears to be much less of a problem in Africa. However, this is not the case,
according to Table 3, which reported higher inequality for the Africa sample. The lower impact of the
Gini coefficient in the regressions is the result of including gross fixed investment. Although the yield
elasticity is slightly reduced, yield is still the most powerful variable in this case.
5.2 Explaining inequality
Thus, the single equation approach is somewhat limited, but this is partly as a result of the
small sample for Africa. However, before presenting some better results for the full sample, we move
on to estimating single equation models to explain inequality. This leads on to a two equation, two-
stage model and a recursive model. The first model reported in Table 18 is a simple experiment to see
if the agricultural productivity variables have a significant impact on inequality. Although their
elasticities are significant and explain 22% of the variance, both are extremely small. Whereas the
correlation coefficients for both value added per unit of land and the land labour ratio were slightly
19
greater than 0.4 in absolute value, in the bivariate correlations of Table 1, they had opposite signs and
labour productivity was not significant, because the two elements cancelled each other out.
Table 18: Dependent Variable is the Gini coefficient: Full sample

Variables
Model 1 Model 2 Model 3 Model 4 Model 5
VA/LAND Negative -0.046* -0.06*** -0.09***
*** **
LAND/LABOUR Negative 0.06 0.96
RURAL POPUPATION (%) Negative -0.20*** 0.13** -0.124***
*** **
POPULATION GROWTH Positive 0.19 0.11
Constant 2.1 4.41 3.53** 4.01** 4.73***
R square 0.22 0.20 0.36 0.18 0.21
F Test 20.3*** 11.73*** 30.3** 13.24*** 16.81***
Sample Size 109 88 102 102 109
*** significant at the 1% level, two-tailed test. ** significant at the 5% level, two-tailed test. * significant at the 10% level,
two-tailed test.
The two other explanatory variables that were significant in these regressions were the
percentage of the population that was rural, which had a negative effect on inequality, and the rate of
population growth, which increased inequality. In model 2, on their own, these two variables
explained less of the variance than did the agricultural productivity variables and the F statistic is far
lower. This model is reported here as a single equation, because it appears later in as one half of a two
equation recursive model, in which the agricultural productivity variables enter only in the other
equation. This route is taken partly because the four variables are not significant if they are included
in a single equation. However, the main reason is that the two should not be estimated independently.
The Gini coefficient is an explanatory variable in the poverty equation and the dependent variable in
the other equation, so the two should be estimated simultaneously. The other three models show the
combinations of the four variables that are significant, but the results are not easy to interpret. For
instance, the rural population variable changes sign when combined with the land labour ratio,
although the other variables are reasonably well behaved. In all, attempts to the attempts to explain
inequality were disappointing.
5.3 Single equations using residuals and recursive systems
Although the Gini regressions shed little light directly, they do suggest that a system of
equations may be the best approach, so these models are developed below. First, model 1 in Table 19
shows that trade as a percentage of GDP can be added to the equation of model 3 in Table 17. It has a
negative effect on poverty, confirming the correlation result of Table 1 that openness is poverty
reducing. However, although the explanatory power is a little better, the F statistic is lower. It does
seem that five variables is as many as the single equation model will accommodate and this is also the
number of variables in the Hanmer and Naschold (2000) model. The choice is somewhat arbitrary: as
model 2 shows, if trade is replaced by the GDP growth rate (not per capita) the model is about equally
acceptable and GDP growth has a poverty elasticity of –0.16. In both these models, the yield elasticity
stays very close to the value of –0.56, that was the estimate in model 3 of Table 17.
Table 19: Dependent Variable is % of Population with Less than $1 per Day
Explanatory Expected Sign
Variables
Simultaneous
VA/LAND Negative -0.55** -0.54** -0.25** -0.49**
LAND/LABOUR Negative -0.67** -0.65** -0.36** -0.64**
GINI Positive 2.28** 1.74** 2.25** 2.33*
20
TRADE (%GDP) Negative -0.26* -0.29*
GROSS FIX INV. Negative -0.80** -0.77 **
-0.84 **
-0.78**
GDP GROWTH Negative -0.16*
GDP PER CAPITA -0.47**
RURAL POPUPATION (%) Negative -0.15**
POPULATION GROWTH Positive 0.24**
Constant 1.73 2.79* 2.44* 1.25
R square 0.59 0.52 0.59 0.51
F Test 18.20** 18.35** 19.3** 83.48**
Sample Size 109 83 107 102
** significant at the 5% level, two-tailed test. * significant at the 10% level, two-tailed test. *** The dependent Variable in the
first part of model 4 is poverty and in the second the GINI Coefficient. The test statistic for the simultaneous model is Chi-
Square rather than F.
This is not true if GDP per capita replaces GDP growth, as it does in model 3. Then, the poverty
elasticity of the yield is reduced to –0.25, which is less than half what it was, and GDP per capita has a
poverty elasticity of –0.47. We would argue that this does not indicate that the previous models were
miss-specified, but rather that higher yields are an important component of higher GDP per capita and
hence the two should not be explanatory variables in the same equation. For comparison, in Hanmer and
Naschold (2000), the poverty elasticity of GDP per capita is –0.93 for the low inequality countries and –
0.34 for those with high inequality. They included the ratio of value added per worker in the modern
sector to value added per worker in agriculture as an explanatory variable. This seems to have a similar
effect to including agricultural productivity as we have done. But, when productivity explicitly competes
with GDP per capita, the yield elasticity was halved, relative to our earlier models.
Further work is required in modelling the relationship if both are to be included, but for now the
problem is ignored and we progress to a recursive two-equation model. Thus, model 4 combines model
1 with model 2 of Table 18, to give equations (5a) and (5b), which are estimated simultaneously using
Zellner’s seemingly unrelated regression model or 3 stage least squares.8 In this model and in all that
follow, all of the variables are in logarithms, so that the coefficients can be interpreted as elasticities.
 VAAG   Land AG 
$1 day poverty = β 0 + β1   + β2   + β 3Gini + β 4Trade + β 5 Fix Invest + ε (5a
 Land AG   LabourAG 
Gini = α 0 Rural Pop + α1 Pop Growth + ν (5b)
In this way the model can be stretched to seven explanatory variables and the results change
what is probably the most theoretically correct model so far, since we conducted tests to ensure this. The
classical assumptions of regression analysis suggest that the standard errors of the fitted coefficients will
be increased if there are correlated independent variables in the model. This may cause the exclusion of
variables that actually have a statistical relation with the dependent variable from the model of study.
Ideally, all variables in a model should be independent; but in practice, it is too strict and may jeopardise
the power of the model. So, the VIF (Variable Inflation Factor) is then used to test the seriousness of the
multicolinearity problem. Many researchers rely on informal rules of thumb applied to the VIF. For
example, Studenmund (1997, p.276) suggests that multicolinearity is severe if the VIF value for any
variables is larger than 5. The land productivity (VA/LAND) and land over labour ratio
(LAND/LABOUR) are correlated to many socio-economic variables. Therefore, in the models we
selected, the VIF value is reported to show that the inclusion of these factors in a model is not violating
the co linearity assumption. That is, the results inferred from these models appear to be reliable, since
Table 20 shows that these variables are not highly correlated.
8
Since the error terms are related, Zellner’s model is required, but 3 stage least squares is not strictly
necessary, as the model is not genuinely simultaneous. The two techniques give very similar results.
21
Table 20: Variable Inflationary Factors
Variable VIF 1/VIF
LAND/LABOUR 2.09 0.477692

VA/LAND 1.85 0.54125
TRADE (%GDP)
1.33 0.749123
GINI 1.3 0.770887
GROSS FIX INV. 1.18 0.84705
Mean VIF 1.55
Both the single equation approach and recursive models have been used in the growth and
poverty alleviation literature. The World Bank poverty and growth studies have frequently applied the
single equation approach, as did Hanmer and Naschold (2000), whereas IFPRI studies, such as Fan,
Hazell and Haque (2000) opt for simultaneous methods. The advantage of the first is apparent
simplicity, but as model 3 in Table 19 showed, there are clear limitations. GDP per capita depends on
so many factors that it is difficult to avoid including the variables that determine it in the single
equation. The effect is apparent when a major determinant of GDP per capita is included, as it was in
this model. There are two ways of overcoming this difficulty and both are used in the next set of
models. The alternative to the recursive equation approach is to resolve the conflict between GDP per
capita and agricultural productivity prior to estimating the single equation.
Thus, model 1, below explicitly deals with the fact that agricultural labour productivity is a
component of GDP per capita. The first equation (6a) has both as independent variables, but the
second equation (6b) regresses GDP per capita on the labour productivity variables. The residual from
this equation may be viewed as GDP per capita net of the contribution of agricultural labour
productivity. Then, it is this agricultural productivity free GDP variable that is used in the first
equation.
Model 1:
$1 poverty = β 0 + β1   + β2   + β 3Gini + β 4Trade + β 5 Fixed Invest
 Land AG   LabourAG  (6a)
+ β 6ε GDP PC + ξ
where GDP per capita = γ 0 + γ 1  +γ2   + ε GDP PC + ζ (6b)
Gini = α 0 + α1Rural Population + α 2 Population Growth + ς (6c)
The effect of this transformation can be seen in the first column of Table 21. The poverty
elasticity of agricultural productivity is now very close to its previous values, at –0.48 and the poverty
elasticity of GDP per capita recovers to –0.68, which is considerable higher than the figure of –0.47 in
model 3 of Table 19. Indeed, this is a little higher than the unweighted average of Hanmer and
Naschold’s (2000) results, which is –0.635. Thus, both agricultural productivity and GDP per capita
can be included in same equation. The third equation (6c) again has the Gini coefficient as the
dependent variable and this is fitted simultaneously with the first equation and produces much the
same results as before. The problem with this model is that the Gini coefficient is insignificant in the
first equation and the reason is clear in model 2.
Table 21: Dependent Variable is % of Population with Less than $1 per Day: Recursive Models
Explanatory Variables Expected Sign Estimated Coefficients
Poverty Reduction
22
VA/LAND Negative -0.48** -0.65**
LAND/LABOUR Negative -0.52** -0.48**
GINI Positive 1.00 1.65** 0.99*
TRADE (%GDP) Negative -0.36** -0.35** -0.34** -0.34**
GROSS FIX INV. Negative -0.68** -0.45** -0.67** -0.66**
GDP PER CAPITA Negative -0.88** -0.76**
AGRICULTURE FREE Negative -0.65** -0.59**
GDP PER CAPITA
AGRICULTURE FREE Positive 2.21**
GINI
CONSTANT 6.01 6.72** 10.40** 8.33**
R-SQUARE 0.52 0.60 0.61 0.52
CHI-SQUARE OR 97.00** 105.61** 16.89** 90.16**
F-STATISTIC
SAMPLE SIZE 100 107 107 100
Gini
RURAL POPULATION Negative -0.18** -0.18**
POPULATION GROWTH Positive 0.23** 0.24**
LAND/LABOUR Negative 0.06**
CONSTANT 4.3** 3.92** 4.30**
R-SQUARE 0.21 0.20 0.21
CHI-SQUARE 27.58** 27.06** 29.61**
SAMPLE SIZE 100 107 100
GDP PER CAPITA

VA/LAND Positive 0.66** 0.65**
LAND/LABOUR Positive 0.68** 0.67**
CONSTANT 3.13** 3.19**
R-SQUARE 0.80 0.78
CHI-SQUARE 423.20** 361.73**
SAMPLE SIZE 107 100
** significant at the 5% level, two-tailed test. * significant at the 10% level, two-tailed test.
Model 2 shows the extent of the inter-relationships, by dropping the agricultural productivity
variables from the first equation (7a). Then, the second column of Table 21 shows that the Gini
coefficient is again significant, as are the percentage of GDP that is traded, gross fixed investment and
GDP per capita. Sixty percent of the variance is explained by these variables. The second equation
(7b) shows that the two labour productivity variables have as much power in explaining inequality as
did the percentage of the population that is rural and the rate of population growth, but in both cases
the R2 remains low at about 0.2. Thus, as the models become more refined, it becomes clear that like
GDP per capita, the Gini coefficient is a function of the labour productivity variables and should not
be included in the same equation. The third equation (7c) uses just the agricultural productivity
variables to explain GDP per capita and Table 21 shows that the explanatory power is amazingly high,
at 80% and so is the Chi-Square test value for this regression.
Model 2:
$1Poverty = β 0 + β1Gini + β 2Trade + β 3 Fixed Invest + β 4GDP per capita + ε (7a)

GDP per capita = γ 0 + γ 1  +γ2   +ξ (7b)
23
Gini = α 0 + α1   + α2   +ζ (7c)
Thus, model 3 is a single equation (8a) in which GDP per capita is the residual, with
agricultural productivity effects removed as in (8b) and with the agricultural productivity effects
similarly removed from the Gini coefficient by using the residual from (8c). The results reported in the
third column of Table 21 show that the poverty elasticity of land productivity now reverts to the
higher level it had in the earlier models (see Table 6), at –0.65. The land labour ratio, percentage of
GDP traded and gross fixed investment are all significant, and so is agricultural productivity free GDP
per capita, which retains a very reasonable elasticity of -0.59. However, this means that the poverty
reducing effects of improving agricultural productivity are actually greater than the effect of
increasing GDP per capita. This seems unlikely, as although agricultural productivity does seem to be
the key variable, it is still only a component of GDP per capita, all be it the main one. Thus, the
broader variable should have the greater impact. Other than this quibble, removing the double
counting of effects caused by the impacts of the productivity variables on both GDP per capita and the
Gini coefficient gives an equation that is entirely satisfactory, that explains 61% of the variance, with
six independent variables. This does seem to be the limit for single equation models of this
relationship, so we now return to recursive models to show how more variables can be incorporated
and explanatory power increased.
Model 3:
$1 poverty = β 0 + β1   + β2   + β 3ε GINI + β 4Trade + β 5 Fixed Invest
 Land AG   LabourAG  (8a)
+ β 6ε GDP PC + ξ
where GDP per capita = γ 0 + γ 1  +γ2   + ε GDP PC (8b)
and Gini = α 0 + α1   + α2   + ε GINI (8c)
It is not really possible to identify one single formulation of the recursive model that
dominates all others, but some of the best models can be briefly covered, to show the range of options.
Model 4 is the same as model 2, except that in equation (9c), the percentage of the population that is
rural and the population growth rate replace the agricultural productivity measures, increasing the
number of variables to eight, instead of six. However, if model 4 is compared with model 2, the
equations fit less well and this is shown by the R2 values and the Chi Square test statistics, which fall
considerably for equations (8a) and (8c).
Model 4:
$1 per day poverty = β 0 + β1Gini + β 2Trade + β 3 Fixed Invest + β 4GDP per capita + ε (9a)
GDP per capita = γ 0 + γ 1  +γ2   +ξ (9b)
Gini = α 0 + α1 Rural Population + α 2 Population Growth + ζ (9c)
Model 5 experiments with the structure of the model by reducing the explanatory variables in
the poverty equation (10a) to just the Gini coefficient and GDP per capita. The, Gini coefficient is
explained by the percentage of the population that is rural and the rate of population growth, in (10 c),
while all the other variables explain GDP per capita growth in (10b). Together, the five exogenous
variables explaining 84% of the variance in GDP per capita, which is reassuring and the model
24
obviously has the highest value for the Chi-Square test for this equation. However, this formulation
fails because the Gini coefficient in (10a) is not significant, as the first column in Table 22 shows.
Model 5:
$1 per day poverty = β 0 + β1Gini + β 2GDP per capita + ε (10a)

GDP per capita = γ 0 + γ 1  +γ2   + γ 3Trade + γ 4 Fixed Invest
 Land AG   LabourAG  (10b)
+ γ 5 Illiteracy + ξ
Thus, model 6 reverts to the previous distribution of the variables between equations and
differs from model 4 only in including labour productivity in industry and services as explanatory
variables in the GDP per capita equation (11b). The results, in the second column of Table 22 show
that neither of these variables is significant, so model 2 remains preferred. Note though, that the test
statistics, such as the Chi-Squares, are reduced not by the effect of having two insignificant variables,
but by the reduction in sample size, due to the number of missing observation for these variables. If it
were possible to construct a larger sample, it is likely that these productivity variables would have a
significant impact, but far lower than for agricultural productivity.
Model 6:
 VAAG   Land AG   VAServices   VAIndustry 
GDP per capita = γ 0 + γ 1  +γ2   + γ3  +γ4   + ξ (11b)
 Land AG   LabourAG   LabourServices   LabourIndustry 
Model 7 differs from 6 only in dropping these two insignificant variables and adding illiteracy
to equation (12b). The third column shows that this is significant and this is probably the best of the
recursive models to date. It has the greatest number of significant explanatory variables (nine): the
highest Chi-Square test statistics, and the best fit overall, in that 51% of the variance is explained in
the poverty equation, 21% in the inequality equation and 82% in the GDP per capita equation. The
values of the two key poverty elasticities are –0.72 for the poverty elasticity of GDP per capita and –
0.40 for the poverty elasticity of land productivity. Since land productivity in the recursive model
affects poverty through its effect on GDP per capita, its elasticity is the calculated as the product, -
(0.72)*(0.55) = -0.40.
Table 22: Dependent Variable is % of Population with Less than $1 per Day: Recursive Models
Explanatory Variables Expected Sign Estimated Coefficients
Poverty Reduction
VA/LAND Negative
LAND/LABOUR Negative
GINI Positive 0.04 -0.216 0.94* 1.74**
TRADE (%GDP) Negative -0.17 -0.37** -0.50**
GROSS FIX INV. Negative -1.16** -0.64** -0.75**
GDP PER CAPITA Negative -0.76** -0.78** -0.72** -0.76**
AGRICULTURE FREE Negative
GDP PER CAPITA
CONSTANT 8.66** 13.92** 8.32** 6.32**
25
R-SQUARE 0.31 0.34 0.51 0.65
CHI-SQUARE OR 45.15** 38.39** 85.24** 85.92**
F-STATISTIC
SAMPLE SIZE 99 42 99 52
GINI
RURAL POPULATION Negative -1.67** -0.08 -0.18** -0.2**
POPULATION GROWTH Positive 0.21** 0.19** 0.23** 0.33**
VA/LAND Negative
LAND/LABOUR Negative
CONSTANT 4.27** 3.97** 4.31** 4.33**
R-SQUARE 0.21 0.21 0.21 0.42
CHI-SQUARE 24.86** 10.54** 28.50** 46.31**
GDP PER CAPITA

VA/LAND Positive 0.55** 0.63** 0.55** 0.58**
LAND/LABOUR Positive 0.62** 0.70** 0.59** 0.57**
VA/LABOUR(SERVICE) Positive 0.12
VA/LABOUR(INDUSTRY) Positive -0.07
TRADE (%GDP) Negative -0.21**
GROSS FIX INV. Positive 0.19**
ILLITERACY Negative -0.16** -0.16**
EXPENDITURE ON Positive -0.14*
PRIMARY EDUCATION
CONSTANT 4.55** 2.76 4.31** 4.06
R-SQUARE 0.84 0.78 0.82 0.77
CHI-SQUARE 517.06** 145.53** 454.26** 177.48**
** significant at the 5% level, two-tailed test. * significant at the 10% level, two-tailed test.
Model 7:
GDP per capita = γ 0 + γ 1  +γ2   + γ 3 Illiteracy + ξ (12b)
The only disappointing result in model 7 is the low explanatory power of the inequality
equation. The last model shows that this is doubled to 42% if illiteracy is replaced by the expenditure
on primary education in equation (13b) and there is the additional benefit that the explanatory power
of the poverty equation increases to 65%, which is as high as it has been in any of the models. The
costs are only a reduction in the R2 for (13b), from 82% to 77% and a reduction in the Chi-Square test
statistic for this regression, balanced by a big improvement in the Chi-Square for the Gini equation,
and the average explanatory power of the three equations is over 60%. The only snag is that the effect
of higher expenditures per student in primary education is negative. This is not so unreasonable, since
the policies may well be donor funded and reactive, so that countries with the most severe problems
26
are spending most. The literacy rate is in many ways a better measure since it records the success of
past education expenditures. Current expenditures have an effect only with a considerable lag, which
has not been allowed for here. Thus, this model is a viable alternative to model 7 and the single
equation results of model 3, in Table 21. The key elasticities are almost identical to the model 7
results, except that the poverty elasticity of the yield rises by ten percent, to –0.44.
Model 8:
GDP per capita = γ 0 + γ 1  +γ2   + γ 3 PEducation + ξ (13b)
Summary
This section has shown that either single equation techniques of recursive equation methods
can explain over 60% of the variance in the percentage of the population living on less than $1 per
day. The two approaches give different answers because the constraints differ. If GDP per capita
and the Gini coefficient are purged of the effects of agricultural productivity growth by prior
regressions, so that all three can be included in the same equation, the poverty elasticity of yield
growth is –0.65. If the yield can only affect poverty through its effects on GDP per capita, as in the
recursive models, then the poverty elasticity is reduced to -0.40, because direct effects are precluded.
Thus, since there are direct effects, we would argue that –0.65 is the more appropriate figure, and is
this value is typical of the results we have produced, however the model has been formulated. Indeed,
this elasticity does seem to be surprisingly robust.
The recursive models do make another point. The question we were asked to investigate was the
effect of agricultural productivity growth on poverty, but it is clear that growth in yields not only affects
poverty, but has an almost equally big impact on GDP per capita, where the elasticity is about 0.58. Thus,
everybody gains from yield growth, not just the poor, so ignoring the GDP per capita gains would be
foolish.
6) R&D COST OF YIELD INCREASES TO REDUCE POVERTY & INCREASE GDP PER
CAPITA
There is a final stage in the development of the recursive model, which is adding a further equation in
which land productivity is explained by agricultural R&D, fertiliser, land quality and illiteracy. The
relationship, which has been extensively modelled (see Thirtle, 1999, for example), should also include
extension expenditures and a weather index, but these data do not exist for most of the sample.9 Thus,
Model 9 adds an additional equation (14d), which completes the model by allowing estimation of the
elasticity of yields with respect to R&D expenditures. The R&D data needs scaling, so it is expressed as
per unit of land, to match up with the dependent variable in this yield equation and fertiliser is treated in
the same way. The land quality index is from the USDA (2001) web site, but it is not included in
equation (14d) because it was not significant.
9
The rationale for this formulation is that R&D generates new technology, which usually requires more
fertiliser. Extension takes the innovation from the trial plot to the farmers and literate farmers fare better at
adapting the technology to their particular circumstances. The weather index accounts for a considerable
proportion of the residual, so the fit would be improved if these data were available.
27
This allows calculation of the marginal internal rate of return to investment in agricultural R&D
for the countries in this sample. Also, once the R&D expenditures for the sample are known, the model
can determine with reasonable accuracy the expenditure necessary to increase GDP per capita by one
percent or to reduce the percentage of the population living on $1 per day by one percent. This allows
calculation of the extra income generated and the number of people who move out of the less than $1 per
day poverty bracket. This final model is specified as Model 9, below, which is fitted to the same sample
as Model 8, since for these countries R&D data is available from Pardey and Roseboom (1989) for all but
four observations.
Model 9:
GDP per capita = γ 0 + γ 1  +γ2   + γ 3 PEducation + ξ (14b)
VA
= φ0 + φ1[ R & D / Land ] + φ2 [ Fertiliser / Land ] + φ3 [ Illiteracy ] + ς (14d)
Land
The results of model 8 are changed only slightly by the additional yield equation. The only
cost is that the inequality equation reverts to explaining only 21% of the variance, but the yield
equation is robust and explains 73% of the variance. R&D, fertiliser and illiteracy are highly
significant and the elasticity of R&D is 0.62. This elasticity is the crucial piece of information needed
to link R&D expenditures to the increase in the value of output that results from the higher yield,
allowing the rate of return to be calculated.
Table 23: Dependent Variable is % of Population with Less than $1 per Day: Recursive Model
Dependent variable, followed by explanatory variables Expected Sign
Poverty Reduction Equation Model 13
% of Population Living on Less than $1 per Day
GINI Positive 1.93**
TRADE (%GDP) Negative -0.57**
GROSS FIX INV. Negative -0.75**
GDP PER CAPITA Negative -0.73**
AGRICULTURE FREE GDP PER CAPITA Negative
CONSTANT 5.59**
R-SQUARE 0.64
CHI-SQUARE OR F-STATISTIC 68.05**
SAMPLE SIZE 48
Inequality Equation
GINI
RURAL POPULATION Negative -0.17**
POPULATION GROWTH Positive 0.22**
28
CONSTANT 4.33**
R-SQUARE 0.21
CHI-SQUARE 13.25**
SAMPLE SIZE 48
GDP per Capita Equation
GDP PER CAPITA
EXPENDITURE ON PRIMARY EDUCATION Positive -0.20**
CONSTANT 4.27**
R-SQUARE 0.80
CHI-SQUARE 167.94**
SAMPLE SIZE 48
Yield Equation
VA/LAND
R&D/LAND Positive 0.62**
FERTILISER/LAND Positive 0.86**
ILLITERACY Negative -0.35**
CONSTANT 14.02**
R-SQUARE 0.73
CHI-SQUARE 164.94**
SAMPLE SIZE 48
6.1 Calculating the Rate of return to Investment in R&D
The rate of return calculation is normally based on the estimated coefficients of R&D in
explaining productivity (Lu, Cline and Quance, 1979, Davis, 1981, Thirtle and Bottomley, 1989, Alston,
Norton and Pardey, 1995), or on the R&D coefficient in the estimation of the dual profit function (Jayne
et al., 1994). The coefficient is an output elasticities relating R&D expenditures to the yield, but it can be
converted to a marginal value products to allow calculation of the marginal internal rate of return (MIRR)
to R&D. If the elasticity is • then
∂LnYieldt ∂Yieldt R & Dt −i

αi = [ ] =[ ][ ]
∂LnR & D t-i ∂R & D t -i Yieldt (15)
where RD and Yield
can be approximated by the mean values, so that the marginal product of R&D in year i is
Yieldt
MP RD t-i = α i [ ] (16)
RD t-i
However, (16) is still in terms of the effect of R&D on yield, and for a rate of return to be calculated, the
change in productivity must be converted into a value. Thus, both sides of equation (16) are multiplied
by a conversion factor that is the change in the value of output (• V), that results from a change in the
yield (• Y). This gives a value marginal product, with both R&D and value added being measured at
constant prices
Yield t ∆V
VMP RD t-i = α i [ ][ ] (17)
RD t-i ∆Y
The marginal internal rate of return (MIRR) can be calculated from equation (18)
29
n
Σ [ (1+ r ) ]
VMPt -i
i
-1 = 0 (18)
i=1
in which, i is the length of the lag, for each expenditure term, and the MIRR for a one unit change in
R&D expenditure is calculated by solving for r. If the lag between R&D expenditures and yield
increases is assumed to be five years, the MIRR is estimated at 52%, which is in line with the
relatively high values normally found in the literature. If the lag were six years, the MIRR falls to
42% and if it were seven years, to 35%.
In this case it is possible to conduct a simple check on the MIRR calculation Data for the
yields of all cereals, for the LDCs, from the FAO Agrostat database allow the rate of growth to be
calculated for the period 1960-1998. The annual average growth rate of yields is 2.4% and Alston et
al. (2000) report that in 1991, the agricultural research expenditures of the LDCs was $8 billion.
Thus, the cost of increasing LDC yields by one percent is $3.3 billion 1991 dollars. This is the cost of
reducing the percentage of the LDC populations living on less than $1 per day by between 0.4% (the
lowest estimate, in model 7) and 0.65% (the single equation result in model 3) and at the same time
increasing LDC GDP per capita by between 0.55% (model 7) and 0.58% Model 8).
Both these effects can be calculated from the estimated elasticities and the data. Alston et al
(2000) do not specify the countries that are included in their group of LDCs, that are spent $8 billion
on R&D in 1991. If we presume they are the same group the FAO puts in this category, their GDP in
1990 was about $2,853.9 billion (from World Bank, 2001). If a 1% increase in yields increases GDP
per capita by 0.565% (the average of the two very similar figures), the gain in GDP for the LDC group
is $16.13 billion and the cost was $3.3 billion.
Whilst this looks attractive, the lags between R&D expenditures and their impact on
productivity are long. The lag is not known and discounting is powerful as Table 24 shows. If the
output gains were in the next the rate of return would be a massive 389%, but if the lag is ten years the
payoff is only 17%. This shows how important it is for LDC national agricultural R&D systems to
concentrate much of their efforts on applied and adaptive research, which has a relatively rapid payoff.
South Africa has been successful in this respect: Khatri and Thirtle (2000) found the peak lag to be at
less than three years and as a result the minimum rate of return estimated was 77%. Most cross
country studies of LDCs, such as Thirtle, Hadley and Townsend (1995) find that a peak lag of about
five years fits the data. If this is about right, the rate of return to agricultural research in the LDCs
should be about 37%, which is entirely satisfactory, and despite the different and somewhat crude
method of calculation, is right in the range that has been found in a large number of studies.
Table 24: Marginal Internal Rate of Return to R&D, by Length of Lag from Expenditure to Benefit
Years of Lag 10 9 8 7 6 5 4 3 2 1
Rate of Return 17% 19% 22% 25% 30% 37% 49% 70% 121% 389%
The MIRR result, from Model 9, of 52% is thus corroborated and this can be also be used to
calculate the gain in GDP per capita that results from a 1% increase in R&D expenditures. In Table
23, the elasticity of GDP per capita with respect to yield is 0.56 and the elasticity of yield with respect
to R&D is 0.62. Thus, the increase in GDP per capita that results from a 1% increase in R&D
expenditures is 0.56 x 0.62 = 0.34%. If this sample is taken to be representative of the LDCs, then the
GDP per capita gain falls from $16.13 billion to 9.7 billion. The rate of return to R&D, calculated
from the GDP gain falls to 24%, but this is still a decent return.
The payoff to agricultural R&D in terms of the number of people that it can shift out of the
less than $1 per day poverty bracket can be similarly calculated. Table 25 first shows the percentages
of the regional populations living on less than $1 per day, followed by the populations. The product
of the two is the number of people living on less than $1 per day. Then if a 1% increase in yields
30
reduces this count by 0.4%, the total number of people escaping from the less than $1 per day bracket
is 4.8 million. If the alternative estimate of 0.65% is used, this total rises to nearly 7.8 million.
Table 25: Reduction in the Number of People on Less than $1 per Day from a 1% Increase in Yield
Region % in $1 Population # in $1 poverty 0.4% 0.65% 0.25%
poverty Reduction Reduction Reduction
East Asia 15.3 1836.9 281.0457 1.124 1.827 0.703

Latin America 15.6 509.2 79.4352 0.318 0.516 0.198
Middle East 1.9 290.9 5.5271 0.022 0.036 0.014
& N Africa
SE Asia 40 1329.3 531.72 2.127 3.456 1.329
SSA 46.3 642.3 297.3849 1.189 1.933 0.698
Total Number of People Moving Out of $1 per Day 4.780 7.768 2.942
Poverty, Millions
Thus, only this small proportion of the estimated (World Bank web site) 1, 200 million people
living on less than $1 per day are removed from this category when $3.3 billion is spent on
agricultural research. This does not sound nearly as promising as the gains in GDP, but we fear it is
correct. To put it in perspective, the models suggest that a 1% increase in GDP per capita has hardly
any more impact than a 1% increase in yields: the model 3 result was 0.59%, which is actually lower
and models 7 and 8 gave 0.72 and 0.76.
The alternative to these simple estimates is to fully use the results of Model 9, which adds
sophistication, but is les direct, as it accumulates the effects of three equations. In Table 23, the
elasticity of $1 per day poverty with respect to GDP per capita is –0.73; the elasticity of GDP per
capita with respect to yield is 0.56 and the elasticity of yield with respect to R&D is 0.62. Thus, the
reduction in the percentage of the population living on less than $1 per day that results from a 1%
increase in R&D expenditures is –0.73 x 0.56 x 0.62 = -0.25%. If this sample is taken b be
representative of the LDCs, then the reduction in the numbers living on less than $1 per day falls to
2.94 million.
These are not spectacular reductions, but the evidence suggests these effects are in the right
range. The data on the World Bank’s Poverty Net web site show that from 1990 to 1998 the number
of people living on less than $1 per day fell from 1,276.4 million to 1,174.9 million. This is a
reduction of 8% over eight years, so the decline is only about 1% per year. Viewed in this way,
agricultural research may well be a useful and cost effective instrument for reducing poverty, but it is a
pretty blunt one as many have argued (Alston et al., 1995). Perhaps getting benefits to the poorest
segment of the population is always the most difficult task.
31
REFERENCES
Alston J, Chan-Kang C, Marra M, Pardey P and Wyatt T (2000) A meta-Analysis of Rates of Return to
Agricultural R&D: Ex Pede Herculem? IFPRI Research Report 113.
Alston, J.M., Norton, G.W. and Pardey, P.G. (1995). Science Under Scarcity: Principles and Practice for
Agricultural Research Evaluation and Priority Setting, Cornell University Press, Ithica.
Bruno, Michael, Martin Ravallion and Lyn Squire (1996). Equity and growth in developing countries:
old and new perspectives on the policy issues. World Bank working paper No 1563
Chen S, Datt G and Ravallion M (1994), Is Poverty Increasing in the Developing World?, Review of
Income and Wealth, 40,(4): 359-76.
Chen S and M Ravallion (2000), How did the poorest fare in the 1990’s, World Bank, Washington, DC.
Davis, J. A. (1981). Comparison of Procedures for Estimating Returns to Research Using Production
Functions, Australian Journal of Agricultural Economics, 25, 60-72.
Fan, S, P Hazell and T Haque (2000), Targeting public investments by agro-ecological zone to achieve
growth and poverty alleviation goals in rural India, Food Policy, 25, 411-28.
Fan S, P Hazell and S Thorat S (1999), Linkages between Government spending, growth, and poverty in
rural India, IFPRI research report 110
Hanmer L and Nashchold F (2000), Attaining the International Development Targets: Will Growth be
Enough?, working paper
Hayami Y and Ruttan V (1986), Agricultural Development: An International Perspective, Johns

Hopkins University Press, Baltimore
Jayne, T.S., Khatri, Y., Thirtle, C. and Reardon, T. (1994). A Profit Function Approach to Productivity
Change in Zimbabwean Agriculture: Implications for Research and Policy, American Journal of
Agricultural Economics, 76, 613-18.
Khatri Y and Thirtle C, (2000). Cointegration and modelling the length and shape of the research lag
Chapter 10 in Colin Thirtle, Johan Van Zyl and Nick Vink (Eds), South African Agriculture at the
Crossroads, Macmillan, Basingstoke.
Lu, Y., Cline, P. and Quance, L. (1979). Prospects for Productivity Growth in US Agriculture,
Agriculture Economic Report, No.435, ESCS, USDA, Washington, DC.
Maxwell S (2001) Innovative and Important, yes, but also Instrumental and Incomplete: the treatment of
redistribution in the new poverty agenda, ODI, London
Pardey, P and Roseboom, H (1989) ISNAR Agricultural Research Indicator Series: A Global Database
on National Agricultural Research Systems, Cambridge University Press
Ravallion M (1997), Can High Inequality Developing Countries Escape Absolute Poverty? Economics
letters 56: 51-57
Ravallion M and G Datt (1999), When is growth pro-poor? Evidence from the diverse experience of
India’s states World Bank Policy Research Working Paper Series, 2263,
32
Studenmund, A. H. (1997), 'Using Econometrics: A Practical Guide', 3rd Edition, New York: Addison-
Wesley.
Thirtle, C., Producer Funding of R&D: Productivity and the Returns to R&D in British Sugar, 1954-93,
Journal of Agricultural Economics, Vol.50, No.3, September, 1999, pp.450-67.
Thirtle, C. and Bottomley, P. (1989). The Rate of Return to Public Sector Agricultural R&D in the UK,
1965-80, Applied Economics, 21, 1063-86.
Thirtle C, Hadley D and Townsend.R (1995). A Multilateral Malmquist Productivity Index Approach to
explaining Agricultural Growth in Sub-Saharan Africa, Development Policy Review, 13: 323-348.
USDA (2001), http://www.ers.usda.gov/Briefing/GlobalResources/Questions/slide4.htm
World Bank (1999), World Development Report 1998/99, World Bank/Oxford University Press
World Bank (2000) World Development Indicators, CD, World Bank, Washington DC.
World Bank (2001), World Development Report 2000/01 Attacking Poverty, World Bank/ Oxford
University Press
33

2334 Background Paper Colin Thirtle Agricultural Productivity Poverty Developing Countries

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

2334 Background Paper Colin Thirtle Agricultural Productivity Poverty Developing Countries

Diunggah oleh

Hak Cipta:

Format Tersedia

1

AGRICULTURAL PRODUCTIVITY AND POVERTY IN DEVELOPING COUNTRIES

Extensions to DFID Report No.7946

Lin Lin*, Victoria McKenzie*, Jenifer Piesse** and Colin Thirtle***

2) METHODOLOGY AND DATA

3) RESULTS: EFFECTS OF AGRICULTURAL PRODUCTIVITY GROWTH ON POVERTY AND NUTRITION

4) DOES PRODUCTIVITY GROWTH IN INDUSTRY AND SERVICES REDUCE POVERTY?

5) EXPLAINING INEQUALITY AND IMPROVING THE SPECIFICATION OF THE MODEL

5.1 Single equation approach

6.1 Calculating the Rate of return to Investment in R&D

2 METHODOLOGY AND DATA

2.1 Basic Regression Models

VALUE ADDED VALUE ADDED LAND

Table 1: Socio-Economic Dataset

Poverty Reducing Correlation Mean Standard Minimum Maximum

All Africa All Africa All Africa All Africa

Explanatory Expected Sign Estimated Coefficients

3.2 Regressions using Pooled Data on $1 per Day Poverty

3.3 Regressions using the Human Development Index

Table 8: Explaining the Human Development Index: Full sample

Table 9: Explaining the Human Development Index: Africa sample

3.4 Regressions using Nutrition Indicators

VA/LAND Positive 0.09**

4 DOES PRODUCTIVITY GROWTH IN INDUSTRY AND SERVICES REDUCE

Table 14: Inter-sectoral productivity correlations

VA/Labour (service) -0.0524 1

VA/Labour (industry) -0.015 0.8879 1

Explanatory Expected Sign Estimated Coefficients

Explanatory Expected Sign Estimated Coefficients

5 EXPLAINING INEQUALITY & IMPROVING THE MODEL SPECIFICATION

5.1 Single equation approach

Table 18: Dependent Variable is the Gini coefficient: Full sample

5.3 Single equations using residuals and recursive systems

Gini = α 0 Rural Pop + α1 Pop Growth + ν (5b)

LAND/LABOUR 2.09 0.477692

Mean VIF 1.55

GDP PER CAPITA

$1Poverty = β 0 + β1Gini + β 2Trade + β 3 Fixed Invest + β 4GDP per capita + ε (7a)

$1 per day poverty = β 0 + β1Gini + β 2GDP per capita + ε (10a)

CONSTANT 8.66** 13.92** 8.32** 6.32**

GDP PER CAPITA

6.1 Calculating the Rate of return to Investment in R&D

∂LnYieldt ∂Yieldt R & Dt −i

East Asia 15.3 1836.9 281.0457 1.124 1.827 0.703

Hayami Y and Ruttan V (1986), Agricultural Development: An International Perspective, Johns

USDA (2001), http://www.ers.usda.gov/Briefing/GlobalResources/Questions/slide4.htm

Anda mungkin juga menyukai

Lin Lin, Victoria McKenzie, Jenifer Piesse and Colin Thirtle*

CONSTANT 8.66 13.92 8.32 6.32